You Get What You Share: Incentives for a Sharing Economy
Total Page:16
File Type:pdf, Size:1020Kb
The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19) You Get What You Share: Incentives for a Sharing Economy Sreenivas Gollapudi Kostas Kollias Debmalya Panigrahi Google Research Google Research Duke University Abstract their individual resources. The platform allows the agents to In recent years, a range of online applications have facilitated list and search for available resources, which enables them resource sharing among users, resulting in a significant in- to identify partners and form sharing groups. More specif- crease in resource utilization. In all such applications, shar- ically, in workforce and educational applications, there are ing one’s resources or skills with other agents increases so- groups that work to complete a task or a project, while in cial welfare. In general, each agent will look for other agents ride sharing and room sharing applications, the notion of a whose available resources complement hers, thereby form- group appears when agents get together to use a provided ing natural sharing groups. In this paper, we study settings resource, e.g., a ride or a house. where a large population self-organizes into sharing groups. A natural goal is to partition the users into sharing groups In many cases, centralized optimization approaches for creat- ing an optimal partition of the user population are infeasible that maximize the overall utility or social welfare of the because either the central authority does not have the neces- system. One may model this as an optimization problem, sary information to compute an optimal partition, or it does where a centralized authority computes the optimal parti- not have the power to enforce a partition. Instead, the central tion of users. However, in typical environments, the informa- authority puts in place an incentive structure in the form of tion necessary for computing this partition is hard to obtain a utility sharing method, before letting the participants form due to privacy issues or due to the sheer size of the popu- the sharing groups by themselves. We first analyze a simple lation. Moreover, the central authority has only limited con- equal-sharing method, which is the one most typically en- trol over the participants, and often cannot enforce the parti- countered in practice and show that it can lead to highly inef- tion. Hence, in this paper, we study the following question: ficient equilibria. We then propose a Shapley-sharing method Consider a population of individuals, who can form shar- and show that it significantly improves overall social welfare. ing groups and share their individual resources with group members. What is the right way to distribute the welfare pro- Introduction duced by each sharing group among its members so that the The sharing economy (Belk 2014) is a term used to describe population self-organizes in a socially optimal manner? The the modern trend of using online platforms and applications semantics of distributing welfare are different for each ap- to increase the value of certain resources (such as goods, ser- plication. For instance, in online learning, utility is given vices, or skills) by enabling the sharing and reuse of these in the form of course credits that depend on how students resources. Carpooling and ride sharing applications (e.g., contributed to group projects. In the ride-sharing applica- Uber, Lyft, RideWith) allow people to better utilize com- tion, using a vehicle has some value that is the same for muting resources (cars, fuel, etc.), lodging and room sharing all agents and monetary transfers reward those providing re- applications (e.g., AirBnB) allow users to share their space sources and extract payments from those utilizing resources when not using it, online learning platforms (e.g., Coursera, to adjust utilities accordingly. Udacity, edX) allow students to share knowledge and col- We model instances of this problem as games that we call lectively learn, and workforce applications (e.g., Upwork, resource sharing games. The most natural way to split the Amazon Mechanical Turk) allow workers to collaboratively produced welfare among the group’s members is to divide it complete requested projects by contributing complementary equally. In fact, this is the method that naturally comes into skills. In all such applications, there are natural limits on the play when there are no monetary or credit transfers in our extent to which the resources can be shared. In carpooling, applications of interest, i.e., everyone simply gets access to each car can fit a small number of people; in lodging, each the resources available in the group. However, we show that room or house has a fixed capacity; in crowsourced projects this method can perform quite sub-optimally with respect to and online courses, there is a prescribed size for working overall social welfare. We then propose an incentive struc- groups. In effect, this means that the agents need to be or- ture derived from the well-known Shapley value in the eco- ganized in groups of limited size so as to effectively share nomics literature (Shapley 1953), and show that the quality Copyright c 2019, Association for the Advancement of Artificial of the resulting outcomes is significantly improved. Intelligence (www.aaai.org). All rights reserved. In our model, each participant owns a subset of all possi- 2004 ble resource types and, when a group is formed, everyone in (in the form of a utility-sharing method) that will lead the the group has access to all resource types that at least one of agents to efficient outcomes. them owns. An important observation is that the resulting so- We begin by studying the most natural and widely used cial welfare function is not submodular, but supermodular. utility sharing method, which dictates that each agent has (This is in contrast to an existing body of work that studies utility equal to the number of resources available in her utility games for submodular welfare functions, e.g., (Vetta sharing group. This is equivalent to equally splitting the so- 2002; Gairing 2009; Marden and Wierman 2013).) Infor- cial welfare produced by a group among its members and, mally, this means that a group produces more utility than hence, we call it equal-sharing (EQUAL). For instance, in the sum of its individuals. While this is often the case in online course projects, EQUAL would mean all students in a practice, it can be theoretically problematic because the only group will get the same course credit, i.e., the problems that stable solution is typically one where all the agents join the the group has solved. In the carpooling application, EQUAL same group. However, in real-world applications (we give means all agents get to take the ride and split any costs some examples below), it is impractical to have very large (e.g., fuel) equally. We show that, for any reasonable equi- groups. So, we will adopt a model where groups are subject librium concept, there is an equilibrium partition of agents to a cardinality constraint, and also show that it is possible into sharing groups that is k times worse than the optimal. to enforce such a constraint through incentives alone. We also explain that this inefficiency with respect to the op- timal is the worst possible for any partition, independent of Applications whether it is an equilibrium or not. In light of this obser- First, we describe two concrete real world examples of our vation, we seek an alternative utility-sharing method, which resource sharing games: results in better equilibria. The method we propose is de- rived from the Shapley value (Shapley 1953) and, hence, we • Massive Open Online Courses – Suppose we are inter- call it Shapley-sharing (SHAPLEY). SHAPLEY aims to im- ested in partitioning the set of students into groups to prove on the main shortcoming of EQUAL, which is that it complete an assignment comprising multiple problems. rewards agents based on their contribution to improving the From a social perspective, we would prefer a partition of social welfare by sharing their resources. Below we present students such that every group in the partition completes our results for the various equilibrium concepts, which show as many problems as possible. Each student on the other that SHAPLEY leads to better equilibria than EQUAL. hand, wishes to maximize the course credit she will re- Nash equilibrium. The Nash equilibrium (NE) is the most ceive. The goal of the central authority (teacher) is to pro- established equilibrium concept in the literature. An NE in vide the right incentives (in the form of a grading scheme) our setting is a partition such that no agent can improve to the students so that effective groups grow organically her utility by unilaterally deviating to a different group. It and in a decentralized fashion. is clear that the set of equilibria of a resource sharing game • Carpooling – In a carpooling system, agents form groups depends on the utility-sharing method being used. The price and repeatedly share rides to and from work during rush of anarchy (POA) (Koutsoupias and Papadimitriou 2009) is hour. In this setting the resources of the game correspond used as a metric that quantifies the performance of a utility- to car rides from location to location and welfare is gen- sharing method by calculating the worst-case ratio of the so- erated for each agent who takes a ride. From a social per- cial welfare in the optimal partition to the social welfare in spective, we wish that the commuting resources are put an NE. As we mentioned earlier, the POA of EQUAL is k. to good use. On the other hand, agents wish to optimize Unfortunately, the POA remains Q(k) for SHAPLEY as well.